City Research Online...Migut and Worring (2010) visualise a classification model, particularly, the...
Transcript of City Research Online...Migut and Worring (2010) visualise a classification model, particularly, the...
-
City, University of London Institutional Repository
Citation: Andrienko, N. and Andrienko, G. (2013). A visual analytics framework for spatio-temporal analysis and modelling. Data Mining and Knowledge Discovery, 27(1), pp. 55-83. doi: 10.1007/s10618-012-0285-7
This is the unspecified version of the paper.
This version of the publication may differ from the final published version.
Permanent repository link: http://openaccess.city.ac.uk/2851/
Link to published version: http://dx.doi.org/10.1007/s10618-012-0285-7
Copyright and reuse: City Research Online aims to make research outputs of City, University of London available to a wider audience. Copyright and Moral Rights remain with the author(s) and/or copyright holders. URLs from City Research Online may be freely distributed and linked to.
City Research Online: http://openaccess.city.ac.uk/ [email protected]
City Research Online
http://openaccess.city.ac.uk/mailto:[email protected]
-
1
A Visual Analytics Framework for Spatio-temporal Analysis and Modelling Natalia Andrienko and Gennady Andrienko
Fraunhofer Institute IAIS (Intelligent Analysis and Information Systems), Sankt Augustin, Germany
Abstract. To support analysis and modelling of large amounts of spatio-temporal data having the
form of spatially referenced time series (TS) of numeric values, we combine interactive visual
techniques with computational methods from machine learning and statistics. Clustering methods
and interactive techniques are used to group TS by similarity. Statistical methods for time series
modelling are then applied to representative TS derived from the groups of similar TS. The
framework includes interactive visual interfaces to a library of modelling methods supporting the
selection of a suitable method, adjustment of model parameters, and evaluation of the models
obtained. The models can be externally stored, communicated, and used for prediction and in
further computational analyses. From the visual analytics perspective, the framework suggests a
way to externalize spatio-temporal patterns emerging in the mind of the analyst as a result of
interactive visual analysis: the patterns are represented in the form of computer-processable and
reusable models. From the statistical analysis perspective, the framework demonstrates how time
series analysis and modelling can be supported by interactive visual interfaces, particularly, in a
case of numerous TS that are hard to analyse individually. From the application perspective, the
framework suggests a way to analyse large numbers of spatial TS with the use of well-established
statistical methods for time series analysis.
Keywords: spatio-temporal data, interactive visual techniques, clustering, time
series analysis
Introduction
It is now widely acknowledged that complex real-world data cannot be properly
and/or efficiently analysed using only automatic computational methods or only
interactive visualizations. Visual analytics research strives at multiplying the
analytical power of both human and computer by finding effective ways to
combine interactive visual techniques with algorithms for computational data
analysis and by developing new methods and procedures where visualization and
computation interplay and complement each other (Keim et al. 2008).
The main role of data visualization is traditionally seen as enabling an analyst to
see patterns in data. Accordingly, visual analytics researchers strive at creating
-
2
techniques capable of effectively exposing various patterns to analyst’s visual
perception. However, the perceived patterns exist only in the analyst’s mind. To
preserve the findings from decay and loss and to be able to communicate them to
others and use in further analyses, the analyst needs to represent the patterns in an
explicit form. Annotation tools, which allow the analyst to supplement visual
displays with text and/or audio notes and drawings, may be sufficient for
supporting recall and communication, but they do not enable the utilization of the
patterns in further computerised analyses. For the latter purpose, the patterns need
to be represented in the form of computer-processable models. Interactive visual
interfaces can effectively support the process of creating such models. Hence,
visual analytics methods and tools should enable not only discovery of patterns in
data but also building of formal models representing the patterns.
Our research focuses on spatio-temporal data, i.e., data with spatial (geographic),
temporal, and thematic (attributive) components. While there are visual analytics
systems supporting the exploration of previously built spatio-temporal models
(e.g. Maciejewski et.al. 2010, 2011), the process of deriving such models from
observed spatio-temporal data has not been yet supported by existing visual
analytics methods and tools. The framework presented in this paper partly fills
this gap. Our approach to spatio-temporal analysis and model derivation can be
briefly described as follows.
Spatio-temporal data often have or can be transformed to the form of numeric
time series (TS) referring to different locations in space or different geographical
objects; such TS will be further referred to as spatial time series, or spatial TS.
Time series analysis and modelling is a well-established area in statistics. The
existing variety of methods and tools can be applied to spatial TS, and we support
this by interactive visual interfaces. However, analysing and modelling each
spatial TS independently from others ignores relationships and similarities that
may exist among spatial locations or objects. To allow these relationships to be
discovered and explicitly represented in the resulting spatio-temporal model, we
employ clustering and interactive grouping, so that related locations or objects can
be analysed together.
The following statements summarise our contribution:
We suggest a new approach to analyse and model spatio-temporal phenomena
described by multiple spatial time series.
-
3
We suggest a comprehensive framework to support the whole process of
analysis and model building. It includes (a) a set of interactive visual tools that
embed existing computational analytical methods and (b) a clearly defined
analytical procedure in which these tools and methods are applied.
In the next section, we give an overview of the related literature. After that, we
introduce our framework for spatio-temporal analysis and modelling and describe
the interactions among the components and the analysis workflow. Then we
present two possible use cases of the framework by example of analysing real
datasets. This is followed by a discussion and conclusion.
Related literature
Linking visual analytics and modelling
There are many works where interactive visualisation is designed to help users to
explore, understand, and evaluate a previously built formal model. Thus, Demšar
et al. (2008) employ coordinated linked views and clustering for exploration of a
geographically weighted regression model of a spatio-temporal phenomenon.
Matković et al. (2010) support users in exploring multiple runs of a simulation
model. Visualisation and interaction can reduce the overall number of simulation
runs by allowing the user to focus on interesting cases (Matković et al. 2011).
Migut and Worring (2010) visualise a classification model, particularly, the
decision boundaries between classes. Interactive techniques allow the user to
update the model for achieving desired performance.
Evaluation of a model often requires testing its sensitivity to parameter values
and/or input data. Visual and interactive techniques are used for exploring the
sensitivity of an artificial neural network model to input data (Therón and De Paz
(2006), the impact of parameter choice on LSA (latent semantic analysis) models
as well as models using the results of the LSA in the further analysis (Crossno et
al. 2009), and the effects of different assumptions on a model of estimated losses
due to a natural disaster, particularly, assumptions about the spatial distribution of
the disaster exposure (Slingsby et al. 2010).
Maciejewski et al. (2010) and Maciejewski et al. (2011) suggest visual analytics
techniques to support the exploration and use of existing spatio-temporal models.
In the former work, kernel density estimation is used for spatial modelling and
-
4
cumulative summation for time series modelling. The user can view and explore
the model results represented on a map and time series display, respectively. In
the latter work, the authors suggest an interactive visual interface allowing the
user to explore the results of a pandemic simulation model and investigate the
impact of various possible decision measures on the course of the pandemic. In
both cases, the user does not participate in model building.
Among the works where interactive visual techniques support the process of
model building, several papers focus on classification models. Xiao et al. (2006)
describe a system visualising network events where the user can select a sequence
of events as an instance of a pattern and describe this pattern by logical predicates
(the system aids the user by suggesting candidate predicates). Then the system
uses this description to find other instances of this pattern in the data. Garg et al.
(2008) suggest a framework where a classifier is built by means of machine
learning methods on the basis of positive and negative examples (patterns)
provided by the user through an interactive visual interface; the user finds the
patterns using visualisations. Garg et al. (2010) describe a procedure in which
clusters of documents are built by combining computational and interactive
techniques, then a classifier for assigning documents to the clusters is
automatically generated, and then the user refines and debugs the model. This is
similar to what is suggested by Andrienko et al. (2009) for analysis of a very large
collection of trajectories: first, clusters of trajectories following similar routes are
defined on the basis of a subset of trajectories, second, a classification model is
built and interactively refined, and, third, the model is used to assign new
trajectories to the clusters.
Visual analytics techniques can also support building of numeric models. Thus,
Guo, Z., et al. (2009) suggest techniques that help an analyst to discover single
and multiple linear trends in multivariate data. Hao et al. (2011) describe an
approach to building peak-preserving models of single time series. However, the
process of deriving spatio-temporal models from multiple spatially referenced
time series has not been addressed yet in the visual analytics literature.
Modelling of spatial time series
Kamarianakis and Prastacos (2006) make a review of methodologies proposed for
spatial time series modelling, particularly, STARIMA, which is a spatio-temporal
-
5
extension of the ARIMA methodology (autoregressive integrated moving
average) widely used in TS analysis. STARIMA expresses each observation at
time t and location i as a weighted linear combination of previous observations
and neighbouring observations lagged in both space and time. This requires prior
specification of a series of weight matrices where the weights define the impacts
among the locations for different temporal lags. The specification of the weights,
which is crucial for the performance of the model, is left to the analyst.
Kamarianakis and Prastacos (2005, 2003) used STARIMA for modelling traffic
flow in a road network based on time series of measurements from 25 locations
and compared it with other approaches applied to the same data. The comparison
showed that a set of univariate ARIMA models built independently for each
location gave better predictions than a single STARIMA model capturing the
entire spatio-temporal variation. The authors attribute this to their simplistic way
of specifying the weight matrices. Although a set of local temporal models (like
ARIMA) is much easier to build and can give better results than a single global
spatio-temporal model, the authors note that excessive computer time may be
needed for building local temporal models in case of hundreds of TS.
We see two disadvantages in modelling by means of STARIMA or similar
methods producing a single global model of the spatio-temporal variation. First,
while the quality of the model critically depends on how well the impacts among
the locations are represented by numeric weights, these impacts may be not fully
clear to the analyst and/or may be hard to quantify. Second, from the user’s
viewpoint, a global spatio-temporal model is a kind of “black box” whose
behaviour is very difficult to understand.
Kyriakidis and Journel (2011) suggest an approach to spatio-temporal modelling
of atmospheric pollution that combines modelling techniques from temporal and
spatial statistics. The temporal variation is modelled independently for each
measurement location and then the spatial variation of the parameters of the
temporal models is, in turn, modelled as a random field. This provides an
opportunity to make predictions for locations for which no measurements are
available. The model is relatively easy to understand for the user since the
temporal models can be explored with the help of time graphs and the spatial
variation of each parameter by means of maps. The approach is applicable to
-
6
spatially smooth phenomena where attribute values change gradually from place
to place.
Generalising from the approaches found in the literature, we can conclude that
spatio-temporal variation can be modelled using methods for TS modelling in
combination with some way of reflecting the spatial variation of the TS. We
suggest clustering of locations or spatial objects by similarity of their TS as a
possible way to represent the spatial variation. It is especially suitable for
representing spatially abrupt phenomena.
An orthogonal approach to decomposing the spatio-temporal modelling task is to
model the spatial variation separately for each time step and then somehow
combine the resulting spatial models to represent also the temporal variation. We
did not find an example of modelling a single spatio-temporal variable in this way
but Demšar et al. (2008) use this idea to model a dependency between several
spatio-temporal variables. Geographically weighted regression (GWR) models are
built separately for several consecutive time steps and then clustering is applied to
the time series of values of the GWR parameters associated with each location. In
this way, the locations are grouped and coloured on a map display according to
the similarity of the respective TS of the parameter values.
Generally, clustering of TS is often used in geovisualisation and visual analytics
for dealing with large numbers of TS.
Visual analysis of multiple (spatial) time series
Ziegler et al. (2010) use clustering to enable visual exploration of a very large
dataset of financial TS. Prior to the clustering, the TS are generalised and
compressed, which not only increases the efficiency of the clustering but also
distils temporal trends from fluctuations. Schreck et al. (2009) apply the Self-
Organising Map (SOM) clustering method (Kohonen 2001) to time series of two
variables and visualise the TS directly within the resulting SOM network layout.
Guo, D. (2009) and Andrienko et al. (2010a,b) apply SOM to spatial TS, assign
colours to the clusters, and use these colours for painting areas in map displays
and lines in time graphs. The time graphs allow the user to see and interpret the
patterns of the temporal variation in the clusters. Andrienko et al. (2010b)
introduce the general idea of representing these patterns by formal statistical
models as a way to externalize results of interactive visual analysis and make
-
7
them practically utilizable. As a proof of concept, the authors made an experiment
on building models for clusters of time series. However, the model building
process was not supported by visual and interactive techniques. Our current work
aims at developing this kind of support.
Presentation of the framework
It is not our goal to develop new methods for TS analysis and modelling since
there are many existing methods and tools, which are widely available in
statistical packages such as R or SAGE or in software libraries such as
OpenForecast or IMSL. Our implementation uses the open-source OpenForecast
library (http://www.stevengould.org/software/openforecast/); however, this should
be considered as just an example.
Components of the framework
The suggested system consists of the following main components:
Cartographic map display, in which spatio-temporal data can be represented
by map animation or by embedded diagrams;
Time series display, shortly called time graph, in which multiple TS can be
represented in summarised and/or detailed way;
Interactive tools for clustering based on one or more of existing clustering
methods, for example, from the Weka library
(www.cs.waikato.ac.nz/ml/weka/) or the SOM Toolbox
(http://www.cis.hut.fi/somtoolbox/);
Methods for TS modelling from a statistical package or library such as
OpenForecast;
An interactive visual interface around the methods from the model library.
Besides these main components, the analysis is supported by tools for interactive
re-grouping (allowing, in particular, modification of computationally produced
clusters), data transformation (e.g., absolute values to relative with respect to the
mean), data filtering (including spatial, temporal, attribute-based, and cluster-
based filters), and display coordination by simultaneous highlighting of
corresponding graphical elements in multiple displays.
The links and interactions among the components are schematically shown in
Figure 1.
-
8
Figure 1. The main components of the framework and links between them.
Analysis workflow
The framework is meant for analysing data having the form of multiple numeric
TS associated with different spatial locations or objects. A time series is a
sequence of values of a numeric attribute referring to consecutive time moments
or intervals, for example, monthly ice cream sales over several years. Each TS is
associated with one location or object in space, for example, a town district or a
café. There are two possible use cases of the framework: (1) analysis of the spatio-
temporal variation of a single space- and time-related attribute, such as the ice
cream sales; (2) analysis of dependencies between two space- and time-related
attributes, for example, sales of ice cream and average air temperature. The two
attributes need to be defined for the same places or objects in space and the same
moments or intervals in time. The analysis workflow is schematically represented
in Figure 2.
Computational tools
ClusteringModelling Data transformations
Data storage Original data Transformed data ClustersModels Residuals
Data selection/ grouping tools
Filtering Direct selection (Re)grouping
Interactive controls
Visual toolsMap display Time series display
User
Legend:Data flow Control flow Perception
-
9
Figure 2. The analysis workflow.
Step 0: Data preparation. When necessary, the analyst transforms the data so as
to make them suitable to the goals of the analysis. For example, the analyst may
divide the amounts of the sold ice cream by the population of the respective
districts and then analyse the ice cream consumption per person.
Step 1: Grouping. The set of TS is divided into groups based on similarity of the
temporal variations of the attribute values. This is supported by the tools for
clustering and re-grouping. The results are controlled using the time graph display
and the map display. The time graph allows the analyst to view the TS of each
group, assess the degree of homogeneity within the group and decide whether it
needs to be subdivided. Two or more groups can be shown in the time graph
together using different colours so that the analyst can compare the groups and
decide whether the differences are high enough or some of the groups should be
united. In the map display, the places or spatial objects characterised by the TS are
painted in the colours of the groups. The analyst can see whether the colours form
meaningful spatial patterns. For example, the places with high ice cream
consumption may be spatially concentrated in the regions with high proportions of
-
10
children and young people in the population. An explainable spatial distribution is
an indication of good grouping.
Step 2: Analysis and modelling. The analyst investigates the temporal variation of
the attribute values in each group of TS by means of the time graph. By visual
inspection, the analyst gains an understanding of the character of the temporal
variation, particularly, whether it is periodic and whether there is a long-term
trend. Thus, ice cream sales may periodically (seasonally) vary reaching higher
values in the summer months and lower values in winter. Besides this seasonal
variation, there may be an overall increasing or decreasing trend.
After gaining an idea in the mind, the analyst externalises it in the form of a curve
that expresses the generic characteristics of the temporal variation within the
group. For this purpose, the interactive interface to the library of modelling
methods is used. The analyst selects the suitable modelling method; for example,
triple exponential smoothing (Holt-Winters method) can be chosen to express
periodic variation with or without a trend. The sequence of input values for the
modelling tool is created, according to the user’s choice, from the median or mean
values taken from all time steps, or from arbitrary percentiles (e.g., 60th
percentiles). When choosing to use the mean values, the analyst may decide to
exclude a certain percentage of the highest and/or lowest values in each time step.
This diminishes the impact of outliers on the mean values.
The generated input sequence, further called representative TS (it represents the
temporal variation in the group), is shown on the time graph and passed to the
modelling tool for building a model. The tool tries to find the best fitting model
(according to statistical criteria such as minimal mean squared error) by varying
the model parameters. After the model is created, the sequence of model-predicted
values for the same time steps as in the original data plus several further time
steps is obtained and shown on the time graph so that the predicted values can be
compared with the representative TS and with the individual TS. The
automatically selected parameters of the model are shown to the analyst.
It is not guaranteed that the automatic selection gives the best possible result.
First, the modelling tool may be trapped in a local optimum. Second, not only the
statistical criteria of fitness are important. Particularly, the model needs to
represent the variation of a group of TS rather than single TS and hence should
have a sufficient degree of generality, which is not achieved by the automatic
-
11
parameter selection. Therefore, the analyst is given the possibility to iteratively
modify the parameters of the model, re-run the modelling tool, and immediately
see the result as a curve on the time graph. When the curve corresponds well to
the analyst’s idea of the general characteristics of the temporal variation in the
group, the model is stored in the data storage.
Step 3: Model evaluation. To evaluate the quality of the model, the analyst
examines the model residuals. The temporal distribution of the residuals is
explored by means of the time graph display and the spatial distribution by means
of the map display. The absence of clear temporal and spatial patterns in the
distribution of the model residuals (in other words, the distributions appearing as
random noise) signifies that the model captures well the general features of the
spatio-temporal variation. If this is not so, the analyst may decide to modify the
model (i.e., return to step 2) or to subdivide the group (i.e., return to step 1) and
refine the analysis.
Step 4. Storing the models externally. Descriptions of the generated models are
stored externally in a human- and machine-readable form such as XML. The
descriptions include all information that is necessary for re-creating the models,
namely: the modelling method, the values of the parameters, and the values
needed for the model initialisation. Besides, the descriptions contain information
about the group membership of the objects or places. The descriptions can be
loaded in another session of the system’s work. The user will be able to view the
models and to use them for prediction. For example, the user may predict the
amounts of the ice cream consumption per person in the next year.
Using the models for prediction
According to our framework, one TS model is built for a group (cluster) of similar
TS associated with different places or spatial objects. If this model were
straightforwardly applied for prediction, the same values would be predicted for
all places/objects of the group and the statistical distribution of the predicted
values would differ from the distribution of the original values. To avoid these
undesired effects, the model-based prediction is individually adjusted for each
object based on the basic statistics (quartiles) of the distribution of its original TS
values. The statistics are computed on the stage of model building and stored in
the model description file together with the information about the group
-
12
membership of the places or objects. Besides that, the statistics of the model-
predicted values for the same time steps as in the original data is computed and
stored together with the description of each model.
The adjustment of the predicted values is done in the following way. Let Q1i, Mi,
and Q3i be the first quartile, median, and third quartile, respectively, of the value
distribution in the original TS for the object/place i. Let Q1, M, and Q3 be the first
quartile, median, and third quartile, respectively, of the distribution of the model-
predicted values for the group containing the object/place i. We introduce level
shift S and two amplitude scale factors Flow and Fhigh as
; ; . Let vt be the model-predicted value for an arbitrary time step t (this value is
common for all group members). The individual value for the object/place i
and time step t is computed according to the formula:
∙ , ∙ ,
The adjustment according to this formula preserves the quartiles of the original
value distribution for each object/place. In the model evaluation step, model
residuals are computed as the differences between the original values and the
individually adjusted predicted values.
Use of the framework
In this section, we describe the analysis workflow in more detail by examples
using two different datasets referring to approximately the same territory (Milan,
Italy). The first dataset, provided by the Italian telecommunication company
WIND, consists of records about 2,956,739 mobile phone calls made during 9
days from 30/10/2008 till 07/11/2008. The second dataset, provided by Comune
di Milano (Municipality of Milan), consists of GPS tracks of 17,241 cars during
one week starting from April 1, 2007. Both datasets have been transformed to
spatial TS by means of spatio-temporal aggregation.
We shall demonstrate two use cases of the visual analytics framework:
1. Analysis and modelling of the spatio-temporal variation of a single space- and
time related variable (by example of the phone calls data).
2. Analysis and modelling of the dependencies between two space- and time-
related variables (by example of the car movement data).
-
13
Use case 1: Analysing the spatio-temporal variation of a single variable
Step 0: Data preparation
For the spatio-temporal aggregation of the phone calls data, we divided the
underlying territory into 307 compartments (cells) by means of Voronoi
tessellation using the positions of the WIND cellular network antennas as the
seeds. Then the call records were aggregated into hourly counts of calls in each
cell, which gave us 307 time series of the length 216 time steps (hours).
Step 1: Grouping
In our example, we use the k-means clustering method from the Weka library, but
other methods can be applied as well. K-means uses the Euclidean distance
between points in the abstract n-dimensional space of attribute values, where n is
the number of the attributes and the points represent the combinations of the
attribute values characterising the objects to be clustered, as the measure of object
dissimilarity. In the case of clustering time series, each time step is treated as a
separate attribute. We run the k-means method with different values of k (number
of clusters) in order to find the most suitable grouping. The results of the
clustering are immediately shown on the time graph (Figure 3) and the map
display (Figure 4) by painting lines in the graph and areas in the map in different
colours assigned to the clusters.
-
14
Figure 3. The lines on the time graph represent the time series of phone call counts clustered by
similarity and painted in the colours of the clusters.
Figure 4. The map display shows two variants of k-means clustering of the mobile phone cells in
Milan according to the time series of the call counts. A: k=7; B: k=9. The legends ((A) and (B))
show the absolute and relative sizes of the clusters.
A good choice of colours for the clusters can facilitate understanding of clustering
results. Our system automatically chooses the colours based on the relative
distances between the cluster centres in the space of attribute values, so that
similarity of cluster colours means closeness of the cluster centres. For this
purpose, the cluster centres are projected on a two-dimensional colour space
(Andrienko et al. 2010a). To allow the user to judge the distances between the
clusters in the attribute space and control the assignment of colours to the clusters,
the system produces a display of the colour space with the projected positions of
the cluster centres (Figure 5). The user can choose one of two suggested schemes
-
15
for mapping between colours and positions on a plane, rectangular or polar. In the
former, four base colours are put in the corners of a rectangle and the colours for
the remaining positions are produced by mixing the base colours proportionally to
the distances from the corners. In the latter, a polar coordinate system is used in
which the colour hue is mapped onto the angular coordinate and the colour
lightness onto the radial coordinate. Figure 5 demonstrates the polar scheme. The
projection of the cluster centres onto the colour space is done using Sammon’s
mapping (Sammon 1969). It is a heuristic iterative algorithm. The user may refine
its result by running additional iterations. The user can also modify the colour
mapping by applying operations ‘flip’ and/or ‘mirror’, which stand for the
symmetric reflection of the colour space along the horizontal and vertical axes,
respectively. In this way, the user can obtain similar colour assignments for sets of
clusters produced in different runs of the clustering algorithm with different
parameters (in our example, different values of k in k-means). This is
demonstrated in Figure 5.
Figure 5. For assigning colours to clusters, the cluster centres are projected on a two-dimensional
colour map. The user can refine the projection and adjust the colours. A: 7 clusters; B: 9 clusters.
On the left is the projection of the cluster centres for k=7 and on the right for k=9.
Obviously, the cluster labels are not the same in the two cluster sets (the labels are
produced automatically from the ordinal numbers of the clusters in the output of
the clustering algorithm). However, the two projections of the cluster centres look
quite similar. They are nearly symmetric with respect to the vertical axis;
therefore, applying the operation ‘mirror’ to one of them results in assigning
similar colours to the cluster centres with similar relative positions among the
other cluster centres of the same set of clusters. In Figure 4 (A, B), these colours
are used for painting the cells on the map. The spatial distributions of the cluster
-
16
colours look very similarly for k=7 (A) and k=9 (B). The colouring of the lines on
the time graph also looks almost the same for k=7 and k=9. This means that
increasing the number of the clusters introduces only minor changes in the
grouping. The projection display and time graph allow us to examine the changes
in more detail. We see that the changes occur in the dark cyan area, where there
are clusters with the lowest values of the call counts (see Figure 3). Increasing k
leads mainly to dividing these clusters into smaller clusters, which do not differ
much from each other (this can be seen, in particular, from the projection of the
cluster centres). Hence, we choose 7 as the reasonable number of clusters.
The map display not only supports comparison of different clustering results but
also allows the analyst to judge the goodness of the grouping based on the
interpretability of the spatial patterns. Unfortunately, we have no local knowledge
of Milan allowing us to interpret the patterns. We can observe that the clusters are
not contiguous in space and that the clusters corresponding to high calling
activities (red, yellow, and light blue) are located at the street ring around the
downtown as well as some of the radial streets connecting the ring to the
periphery. We have also compared the distribution of the clusters to the Milan
map of metro and tram lines, which could be found in the Internet, and found that
many of the hot spots of the calling activities are located at crossings of two or
more transportation lines. Hence, the observed spatial patterns can be partly
related to the transportation network topology. We can speculate that the patterns
can also be related to the areas of business activities in Milan, but we have no
information for checking this.
After testing the impact of the clustering parameter and choosing the suitable
value, we review the resulting clusters one by one to judge their internal
homogeneity. If some cluster has high internal variability, it should be subdivided
into smaller clusters. We do this by means of progressive clustering (Rinzivillo et
al. 2008), i.e., applying the clustering tool to members of one or a few chosen
clusters. This is illustrated in Figure 6, where the time series are shown on a time
graph display in a summarised way, as described in (Andrienko and Andrienko
2005). Instead of representing the individual time series by lines, the display
shows the frequency distribution of the values in each time step by polygonal
stripes shaded in alternating light and dark grey. The polygon boundaries are built
by connecting the positions of the corresponding quantiles (e.g., quartiles,
-
17
quintiles, or deciles, according to the user’s choice) in consecutive time steps. The
screenshots in Figure 6 show the quintiles. The thick black line connects the
positions of the average values.
Figure 6. A selected cluster with high internal variability (cluster 6) has been subdivided into 3
clusters by means of progressive clustering. A: The time series of the original cluster in a
summarized form. B, C, D: The time series of the resulting three clusters (clusters 6, 8, and 9).
Figure 6A summarises the 32 time series of cluster 6. The internal variability is
quite high, as may be judged from the widths of the inner quintile stripes (the
outer stripes are less important for assessing the variability as they may contain
outliers). It is especially high on Saturday and Sunday (days 3 and 4; the vertical
lines on the graph separate the days). We use k-means to divide cluster 6 in two,
three, and four smaller clusters and find that the division into three clusters is the
most reasonable. With two clusters, we still have high variability in one of them,
and with four clusters, we get two very small clusters (with 4 and 5 members)
while the variability is not noticeably reduced. The sections B, C, and D of Figure
6 show the TS of the three clusters obtained by refining cluster 6. It can be seen
that the clusters mainly differ in the values attained on Saturday and Sunday. The
division has substantially reduced the inter-cluster variability, particularly, among
-
18
the values for the weekend. The spatial pattern on the map has remained almost
the same since the colours assigned to the new clusters are very close to the
original colour of cluster 6.
As a summary of the grouping results, Figure 7 shows the time series of the mean
values of the final nine clusters.
Figure 7. The temporal variation of the mean values in 9 clusters of cells.
Figure 8. An interactive visual interface for the temporal analysis and modelling: A) Check
automatically detected time cycles in the data. B) Select the current class (cluster) for the analysis
and modelling. C) Build the representative TS. D) Select the modelling method. E) View and
modify model parameters (this section changes depending on the selected modelling method).
Step 2: Analysis and modelling
An interactive visual interface for the temporal analysis and modelling is shown in
Figure 8. We shall not describe this specific UI in detail since the UI design is not
the focus of the paper. It is more important what the user can and is supposed to
do, according to the suggested framework. The user considers the previously
defined clusters of TS one by one; this is supported by controls for cluster
selection. The TS of the selected cluster can be seen in a detailed or summarized
way in a time graph. The representative TS for the modelling is built from the
-
19
mean or percentile values, as described earlier, according to user’s choice. The
user selects one of the available modelling methods from the library. Depending
on the selected method, a set of controls for specifying model parameters appears.
The user can limit the time range of the values to be used for deriving the model.
In Figure 8, the beginning and end of the selected time range are marked by green
and red vertical lines, respectively.
The user can run the modelling tool without specifying values for the model
parameters. The modelling tool will try different values or value combinations, in
case of two or more parameters, to come to the best fitting model. The resulting
model is represented in the time graph by a curve (as the yellow-coloured curve in
Figure 8). To build the curve, the model is used to predict the values for the time
steps originally present in the data plus several further time steps. It is possible to
see if the model captures well the shape of the input curve and if the prediction for
the further time steps is plausible. The automatic procedure for finding the best
fitting model does not necessarily produce a good result. The current model
parameters are shown to the user, who can modify them and run the model-
building tool repeatedly until the result is satisfactory. Thus, the model presented
in Figure 9 has been obtained after some modifications of the smoothing
parameters. It fits much better the input data than the one in Figure 8.
Figure 9. A result of modifying model parameters.
Some modelling methods, such as triple exponential smoothing (Holt-Winters
method), assume periodic (cyclic) variation of the data. For building a model, it is
necessary to specify the length of the cycle, i.e., the number of time steps. To help
the user to deal with cyclic variation in the data, the tool tries to detect
automatically which of the common temporal cycles (daily, weekly, and yearly)
are present within the time span of the data. The detected cycles are described in
the UI of the tool, as in Figure 8A. The user is expected to check whether the
cycles have been identified correctly. The user can also specify which step in the
data must be treated as the start of each cycle. When the user selects a modelling
-
20
method assuming cyclic variation in the input data, the tool automatically fills the
required information about the cycle length and cycle start in the relevant fields
for specifying model parameters (Figure 8E). The user can modify these
automatically filled values if needed.
The data may involve more than one temporal cycle. This is the case in our
example dataset, where the values vary according to the daily and weekly cycles.
As can be seen in Figure 8, the tool has automatically detected the daily cycle
consisting of 24 time steps of one hour length and the weekly cycle consisting of
168 time steps (i.e., 7 times 24 hours). The time period of the data starts from 0
o’clock on Thursday but we want to use 0 o’clock on Monday as the beginning of
the weekly cycle. Therefore, we specify that the weekly cycle starts from step 96.
To our knowledge, there is no time series modelling method that can deal with
two or more time cycles, or, at least, there is no such method in the openly
available libraries we have investigated. In our tool, we have enabled two
approaches that can be used for data with two cycles:
a) Ignore the larger (outer) cycle and build a model assuming that the data varies
only according to the smaller (inner) cycle. In our example, we would ignore
the weekly variation and consider only the daily variation.
b) Build a combination of models with a separate model for each position of the
smaller cycle within the larger cycle. Thus, for the case of the daily and
weekly cycle, separate models are built for Mondays, Tuesdays, ..., Sundays,
i.e., the variation is represented by a combination of seven models.
Figures 8 and 9 correspond to approach a. Figure 10 shows a model obtained with
approach b. The latter model is better in representing the differences between the
working days and weekend and in giving plausible predictions for the future.
Figure 10. A representation of the daily and weekly variations by a combination of models with a
separate model for each day of the week.
-
21
One note should be made here. For building a model representing cyclic variation
in the data, it is necessary to have an input TS with at least two full cycles. For
example, to build a model for each day of the week, we need data from at least
two Mondays, two Tuesdays, and so on. If the time span of the available data is
shorter, our tool allows the user to construct a longer input TS for the modelling
method either by doubling the representative TS of the group or by concatenating
several specific TS selected from the group. In the latter case, the tool selects the
TS that have the closest values to the representative TS. The number of the
specific TS to use is chosen by the user.
Step 3: Model evaluation
Besides the visual inspection of the curve representing the result of the modelling,
the quality of the model is assessed by analysing the residuals, i.e., the differences
between the real values and the model-predicted values. The tool automatically
computes the residuals for each of the original TS.
In evaluating a model, the absolute values of the residuals are not important, i.e.,
high residual values do not necessarily mean low model quality. The model needs
to capture the characteristic features of the temporal variation but not reproduce
all fluctuations present in the original data. The residuals are expected to reflect
the fluctuations. The goal is not to minimise the residuals but to have them
randomly distributed over time, which means that the model captures well the
characteristic, non-random features of the temporal variation.
To see how the residuals are distributed over time, we look at the time graph
representing the time series of the residuals for the cluster that is currently under
analysis. Representing the individual TS as lines results in a highly cluttered
display; therefore, we look at summarised representations. In Figure 11, the TS of
the residuals of one cluster (cluster 3) are summarised in two ways. The upper
part of the graph shows the quintiles of the frequency distributions. The lower part
is a temporal histogram with a bar for each time step divided into segments
proportionally to the counts of residual values in user-specified intervals. The
segments are painted in colours assigned to the intervals.
-
22
Figure 11. The model residuals for one cluster of TS (cluster 3) are summarised in a time graph
display in two different ways.
Both summarised representations show us that there is some non-random feature
in the residual values: in the evenings of the working days negative values are
more frequent than positive values. To find the reason for this undesired feature,
we need to take a closer look at the TS of the residuals. To reduce the workload,
we do not analyse each individual TS but group them by similarity, as we did for
the original TS, and look at the groups. Figure 12 shows two of the groups. The
thick black lines connect the average values in the consecutive time steps.
Figure 12. Two groups of TS of model residuals.
In the upper graph, the variation of the residuals appears to be random; no regular
features are observed. In the lower graph, we can observe periodic drops of values
occurring in the evenings. This means that the generic model built for cluster 3
-
23
does not capture well enough the pattern of the temporal variation in the
corresponding subgroup of cells. More specifically, the model consistently
overestimates the evening call counts in these cells. To improve the model quality,
we need to subdivide the cluster so that the cells with non-random residuals are
separated from those with random residuals and then build new, more accurate
temporal models for the resulting clusters. The division can be done again by
means of clustering or simply by interactive grouping (classification). In our case,
we subdivide cluster 3 into two clusters according to the residuals. After building
models for these clusters, we find that the undesired regular pattern has been
eliminated.
After the modelling is done for all groups of cells, the descriptions of the models
are stored externally together with information about the group membership of the
cells, the statistics (minimum, maximum, and quartiles) of the distribution of the
original values for the cells, and the statistics (mean and standard deviation) of
their residuals. At any time, the descriptions can be loaded in the system and the
models displayed in a graphical form, i.e., as curves representing the generic
features of the temporal variation in the groups.
Use of the models
The descriptions of the models and objects can be used for predicting values for
new time steps that were not present in the original data. When there is no
periodic pattern in the data, only short-term predictions for the next few time steps
after the end of the original TS can be made. When temporal variation is periodic,
as in our case, and it is assumed that the pattern does not change over time, it is
possible to make predictions for times farther in the future with respect to the
times of the original data. To obtain a prediction, the analyst needs to specify the
time interval for which the prediction will be made. When the prediction is
computed, the system can, depending on the user’s choice, introduce random
(Gaussian) noise in the predicted values according to the statistics of the model
residuals for each object. As an example, the time graphs in Figure 15 show the
call counts for the period from 09/01/2012 (Monday), 0 o’clock to 15/01/2012
(Sunday), 23 o’clock predicted by the models that we have built. The upper and
lower graph (A, B) represent the predicted values without and with the noise,
respectively. The patterns of the predicted values match very well the patterns of
-
24
the original values. Note that the predicted TS start on Monday unlike the original
TS starting on Thursday (Figure 3). The statistics (means and quartiles) of the
distribution of the predicted values coincide with those for the original values.
Figure 13. The time graph shows the predicted call counts for the period from 09/01/2012
(Monday), 0 o’clock to 15/01/2012 (Sunday), 23 o’clock. A: The prediction without introducing
noise. B: The prediction with introduced Gaussian noise.
Figure 14. The use of the model for detecting anomalies. Black: two cells at the San Siro stadium.
The predictions given by the models (and by this class of models in general) can
be used for different purposes, such as optimising network topology or planning
maintenance works. One of the possible uses is detection of anomalies (very high
deviations from expected values) in current or historical data. For example,
anomalies in the call counts are easily seen in Figure 14, which represents the
differences between the real and predicted values, i.e., the model residuals. The
most extreme deviations from the predicted values occur in two cells: on the
working days, in the cell including piazza Ovidio on the east-southeast of the city
(increases up to 1139 over the predicted values, often at 14 and 19 o’clock), and
on Saturday and Sunday in the area of Ospedale San Raffaele on the east-
-
25
northeast (increases up to 848 at 10 o’clock on Sunday). We have no local
knowledge or additional information sources to explain these anomalies.
However, by searching in the Internet, we could explain smaller anomalies
occurring in two cells at the San Siro stadium. The corresponding lines are
highlighted in black in Figure 14. The deviations of the real values from the
predicted ones are between 140 and 330 in the evening of 02/11/2008 (Sunday),
especially at 19:00 (330) and 22:00 (262). This corresponds to a football game
attended by about 50,000 spectators. A smaller increase of values in these cells is
observed in the evening of 06/11/2008 (Thursday); the deviations from the
predicted values range between 49 and 92. This corresponds to another football
game attended by 11,000 spectators.
Use case 2: Analysing the dependence between two spatio-temporal variables
Step 0: data preparation
In this example, we use the GPS tracks of the cars in Milan. First, we did spatio-
temporal aggregation of the data using the method by Andrienko and Andrienko
(2011). The territory of Milan was divided into spatial compartments (cells) and
the time span of the data (one week) was divided into 168 hourly intervals. For
each ordered pair of neighbouring cells and each time interval, the aggregation
tool computed the number of cars that moved from the first to the second cell as
well as the average speed of the movement. The resulting TS are associated with
spatial objects called ‘aggregate moves’, or ‘flows’, which were created by the
aggregation tool. A flow is a vector in space defined by a pair of locations (points
or areas), the start location and the end location. In our example, there are 2155
flows. The counts of the moving objects (cars in our case) are often called ‘flow
magnitudes’. In a map display, the flows are represented by special ‘half-arrow’
symbols, which can show movements in two opposite directions (Figure 15).
Attributes of the flows, such as magnitudes or average speeds, can be represented
by varying the thickness of the lines or by colour coding. When we compare the
time graphs representing the TS of flow magnitudes (Figure 16 top) and average
speeds (Figure 16 bottom), we see that the speeds tend to decrease at the times
when the flow magnitudes increase. We aim to capture this relationship by
numeric dependency models.
-
26
Figure 15. Flows between cells are represented by directed symbols (half-arrows) with the widths
proportional to the total counts of objects that moved. For a better display legibility, minor flows
(with counts less than 150) have been hidden. Left: the whole territory; right: the northern part
enlarged.
Figure 16. The temporal variation of the flow magnitudes (top) and average speeds (bottom) are
shown in a summarised form. The graphs represent the deciles of the frequency distributions
(stripes in light and dark grey) and the mean values (thick black lines).
For this modelling task, we do an additional transformation of the data. The
system allows the user to transform two time-dependent attributes A and B
defined for the same time steps into series of values of B corresponding to
different value intervals of A. For this purpose, the user divides the value range of
attribute A into suitable intervals. For each interval and each object/place, the
system finds all time steps in which the values of A belong to this interval and
collects the values of B attained in these time steps. From the collected values of
B, the system finds the minimum, maximum, mean, and quartiles. In this way, the
system derives a family of attributes: minimum of B, mean of B, first quartile of
B, median of B, and so on. For each of the derived attributes and each object,
-
27
there is a sequence of values corresponding to the chosen value intervals of
attribute A. These sequences are similar to time series except that the steps are
based not on time but on values of attribute A. We shall call these sequences
dependency series (DS) since they are meant to express the dependency between
attributes A and B. In this transformation, attribute A is treated as the independent
variable and B as the dependent variable.
In our example, we take the flow magnitude as the independent variable and the
average speed as the dependent variable. The values of the flow magnitude range
from 0 to 69. We divide this range into intervals of length 3: 0-2, 3-5, 6-8, and so
on; 23 intervals in total. From the attributes the system has computed, we take the
attribute “Maximum of average speed” in order to analyse what speed can be
potentially reached depending on the flow magnitude. Before the analysis, we
filter out the flows where the maximum magnitude is below 5, which means that
there is not enough data for identifying any dependency. We shall deal with 1186
flows out of the original 2155.
Step 1: Grouping
We group the dependency series of the attribute “Maximum of average speed” by
similarity using one of the available clustering tools. The clustering is applied
only to the objects that satisfy the current filter. Figure 17 shows the results of the
grouping. We have clear and easily interpretable spatial patterns, which
correspond to the street network topology of the city. To visualise the DS, we use
a line chart display, which is the same as a time graph display but the horizontal
axis represents an independent variable (in our example flow magnitude) rather
than time. The DS are represented in the same way as TS.
-
28
Figure 17. The flows have been clustered according to the similarity of the dependency series of
the attribute “Maximum of average speed”. Left: the spatial distribution of the clusters. Right: the
DS are represented by lines on a line chart.
Step 2: Analysis and modelling
Modelling of dependencies is done almost in the same way as modelling of
temporal variations except that temporal cycles are not involved. The analyst is
expected to select the range of the values of the independent variable that will be
used for building the model. This needs to be done separately for each cluster (in
the previous use case the same time range was suitable for all clusters). As can be
seen from Figure 17 (right), not all lines representing the DS have the same
horizontal extent. This is because there are many flows where high values of the
independent variable are not reached and, hence, there are no corresponding
values of the dependent variable. Therefore, to build a dependency model for a
group, the analyst needs to select a subsequence of values of the independent
variable for which there are enough values of the dependent variable. An
additional reason for limiting the value range for model building is the reliability
of the data. Thus, in our example, the first value interval of the flow magnitude is
from 0 to 2. The corresponding average speeds have been computed from
movements of at most two cars; hence, the values cannot be sufficiently reliable.
It may be reasonable to ignore these values in the course of dependency analysis
and modelling, i.e., exclude the first value interval of the independent variable.
Figure 18 gives two examples of limiting the range of the independent variable for
model building. The green and red vertical lines mark the beginning and end of
the selected sub-range, respectively. The user can interactively move the limits
and try to build dependency models on the basis of different selections.
-
29
Figure 18. To build a dependency model for a group of DS, the analyst needs to select a valid
subsequence of values of the independent variable. The beginning and end of the subsequence are
marked by green and red vertical lines, respectively.
There are modelling methods that are only applicable to time series and cannot be
used to model dependencies. In the OpenForecast library, the methods suitable for
dependency modelling are linear regression and polynomial regression. When
polynomial regression is chosen, the analyst needs to specify the order of the
polynomial that will be generated.
Step 3: Model evaluation
The evaluation of the models is done in the same way as in the case of TS
modelling, i.e., by exploring the distribution of the model residuals. Here we
briefly present one example. When we look at the statistical distribution of the
residuals for the group of DS shown on the top of Figure 18, we notice that the
model mostly overestimates the values of the dependent variable. This can be seen
in the temporal histogram in Figure 19D: in almost all time steps there are much
more negative residual values (represented by shades of blue) than positive ones
(shades of red). We try to improve the model by changing the representative DS
-
30
and the range limits of the independent variable. The residuals are immediately re-
computed and the display of the residuals is updated as soon as the model
changes. By observing the display, we see that our attempts do not bring
sufficiently good results. We decide to refine the model by dividing the group into
smaller groups, as shown in Figure 19A-C. Figure 19E shows the temporal
histogram of the residuals after the refinement. The sizes of the yellow segments,
which represent residual values close to zero (more precisely, from -2 to +2), have
increased, and the prevalence of the blue segments (negative values) has been
removed.
Figure 19. A,B,C: Based on residual analysis, a group of DS has been divided into 3 subgroups,
for which separate dependency models have been built. D, E: The temporal histograms of the
distribution of the residuals before (D) and after (E) the division and model refinement.
Use of the models
In Figure 20, the dependency models that have been built are represented as
curves on a line chart. This representation of the models can be obtained at any
time after loading the stored description of the model set in the visual analytics
system. Hence, the models can be reviewed by the creator and communicated to
others.
Figure 20. The dependency models are represented as curves on a line chart. The colours of the
curves correspond to the clusters presented in Figure 17.
-
31
Figure 21. The maximal average speeds of the flows predicted on the basis of the actual flow
magnitudes.
Figure 21 demonstrates the use of the dependency models for prediction. By
means of the models, the maximal average speeds have been predicted based on
the actual flow magnitudes by hours in the original TS. The predicted values also
form time series defined for the same sequence of time steps as the original TS.
As can be seen, the character of the temporal variation of the predicted values is
the same as in the original TS of the average speeds (Figure 16 bottom) while the
fluctuations have been reduced. The predicted values are generally higher than the
original values. This is explainable since the modelling has been built on the basis
of the maximal average speeds. We have also repeated the model building
experiment for the means and medians of the average speeds. The resulting
models, when applied to the original TS of flow magnitudes, also convey well the
character of the temporal variation; however, the predicted values are lower than
in the original TS of average speeds. The choice of the suitable attributes for
dependency modelling may depend on the analyst’s goal. Thus, the dependency
models based on the maximal average speeds can be used for estimating the
required travel time for a given route depending on the current or predicted traffic
conditions.
Discussion
Our framework models spatio-temporal variation through grouping locations or
spatial objects by similarity of their TS. The cartographic display and interactive
tools allow the analyst to define groups that form meaningful spatial patterns. If
necessary, the groups may be built in such a way that only neighbouring locations
or objects are put together; e.g., Guo, D. (2009) suggests several algorithms for
obtaining spatially contiguous clusters. Since any way of grouping divides the
territory into regions with crisp boundaries, the grouping-based approach is
suitable for modelling spatially abrupt phenomena, i.e., such that neighbouring
-
32
places or objects can substantially differ by their characteristics. Phone calls and
traffic flows are examples of such phenomena. The same character of the spatial
variation can also be observed for other phenomena related to human life and
activities, for example, voting during elections or housing prices.
To model spatially smooth phenomena, such as atmospheric pollution (Kyriakidis
and Journel 2001), it is valid to apply modelling methods that involve spatial
smoothing of the available data, which is not supported by our framework. On the
other hand, it is not valid to apply models involving spatial smoothing to spatially
abrupt phenomena. Most of the available methods for spatial modelling assume
smooth spatial variation. Dealing with spatially abrupt phenomena requires other
approaches, and clustering is one of the possibilities. Our framework includes
flexible and easily steerable tools for clustering that allow the user to represent the
spatial variation in accord with user’s understanding of the data and background
knowledge of the underlying phenomena.
The temporal variation, according to our framework, is modelled by means of
existing methods for time series modelling. This field of statistics is well
developed, and the methods are widely available and commonly used. We do not
limit the framework to a particular set of methods present in this or that library or
package but propose a generic set of interactive controls and operations suitable
for different methods. In our paper, we have demonstrated how temporal models
and dependency models can be built using these controls and operations.
A very important step in model building is model evaluation. Our framework
supports this step by (a) immediate visualisation of model results (predicted
values) as soon as a model is built, (b) immediate computation and visualisation
of model residuals, and (c) providing possibilities for applying various analysis
methods to the residuals in order to examine their spatial and temporal
distributions. Besides interactive analysis, as demonstrated in this paper, it is
possible to compute and analyse distribution statistics and apply the available TS
analysis methods to the TS of residuals. In case of unsatisfactory residuals, the
analyst can easily find the reason and the way to improving the model.
We do not claim that our framework guarantees better quality of generated models
than other tools. The quality of a model can only be ensured by diligent work of
the analyst. However, the flexibility in choosing modelling methods and
parameters and the support for model evaluation and refinement help a diligent
-
33
analyst to build high quality models. The framework also allows the analyst to
control the degree of data abstraction and generalisation and achieve a suitable
trade-off between the model quality and model complexity, i.e., the number of
different statistical models that represent the entire spatio-temporal variation.
Kamarianakis and Prastacos (2003) compared several methods for modelling
spatio-temporal data and found that a global spatio-temporal model does not
necessarily perform better than a set of local temporal models. One of the
arguments in favour of a global model was the excessive computational time
needed for building multiple local models in case of a very large dataset.
Clustering makes our approach suitable for large collections of spatial TS. It
reduces not only the required computational time but also the analyst’s effort as
compared to building either a global spatio-temporal model (which requires
tedious specification of weight matrices) or individual local models for all
locations. Besides, clustering is a way to account for spatial relatedness among
locations or objects and to represent patterns of spatial variation.
The suggested approach is sufficiently generic to be applicable to various spatio-
temporal phenomena that can be characterised by spatial time series, such as
sensor measurements, economical data, aggregated spatial events, aggregated
movements, and many others. There is a potential for extending the approach to
spatially smooth phenomena by involving spatial modelling methods.
Conclusion
The suggested framework is based on visual analytics wrapping around
computational methods taken from the areas of machine learning and statistical
analysis. The framework is intended for analysis and modelling of spatio-temporal
data representing spatially abrupt dynamic phenomena. The result of the analysis
is an explicit parsimonious description of the spatio-temporal variation, i.e., a
description involving data abstraction and generalisation. The result is represented
in a form allowing both human and computer processing. The framework is
designed to deal with large amounts of spatial TS that cannot be analysed
individually.
From the visual analytics perspective, our framework suggests a way to represent
results of interactive visual analysis in an explicit form, which can be not only re-
viewed and communicated but also used in further analysis and for making
-
34
predictions. From the statistical analysis and modelling perspective, we suggest a
combination of visual, interactive and computational techniques supporting model
building and evaluation. From the perspective of spatio-temporal analysis, we
suggest an approach to spatio-temporal modelling by decomposing the overall
modelling task into spatial and temporal modelling subtasks. We accommodate
the statistical methods for time series and dependency modelling, which are well
established and widely available, for spatio-temporal analysis. For this purpose,
these methods are combined with interactive clustering and visual techniques.
As noted by our colleagues practicing statistical analysis of various types of data,
our visual analytics approach to modelling increases analyst’s trust in the resulting
models as the analyst fully controls the process and can make well-informed
decisions in each its step.
Acknowledgments
This work was supported by the European Commission within the international
cooperation project ESS – Emergency Support System (contract No 217951) and
by DFG – Deutsche Forschungsgemeinschaft (German Research Foundation)
within the Priority Research Programme "Scalable Visual Analytics" (SPP 1335).
We are thankful to our colleagues who discussed with us our work and gave us
helpful comments.
References
Andrienko, G. L., Andrienko, N. V. 2005. Visual exploration of the spatial distribution of temporal
behaviours. In 9th International Conference on Information Visualisation IV2005, 6-8 July 2005,
London, UK, IEEE Computer Society, pp. 799–806.
Andrienko, G., Andrienko, N., Rinzivillo, S., Nanni, M., Pedreschi, D., Giannotti, F. 2009.
Interactive Visual Clustering of Large Collections of Trajectories. In Proc. IEEE Symposium on
Visual Analytics Science and Technology VAST’09, pp. 3-10
Andrienko, G., Andrienko, N., Bremm, S., Schreck, T., von Landesberger, T., Bak, P., Keim, D.
2010a. Space-in-Time and Time-in-Space Self-Organizing Maps for Exploring Spatiotemporal
Patterns. Computer Graphics Forum, 29(3), pp. 913-922.
Andrienko, G., Andrienko, N., Bak, P., Bremm, S., Keim, D., von Landesberger, T., Pölitz, C.,
Schreck, T. 2010b. A Framework for Using Self-Organizing Maps to Analyze Spatio-Temporal
Patterns, Exemplified by Analysis of Mobile Phone Usage. Journal of Location Based Services,
4(3/4), pp. 200-221
-
35
Andrienko, N., Andrienko, G. 2011. Spatial generalization and aggregation of massive movement
data. IEEE Transactions on Visualization and Computer Graphics, 17(2), pp.205-219
Crossno, P.J., Dunlavy, D.M., Shead, T.M. 2009. LSAView: A tool for visual exploration of
latent semantic modelling. In Proc. IEEE Symposium on Visual Analytics Science and
Technology VAST’09, pp. 83-90
Demšar, U., Fotheringham A.S., Charlton M. 2008. Exploring the spatio-temporal dynamics of
geographical processes with Geographically Weighted Regression and Geovisual Analytics.
Information Visualization, 7, pp.181-197
Garg, S., Nam, J.E., Ramakrishnan, I.V., Mueller, K. 2008. Model-driven Visual Analytics. In
Proc. IEEE Symposium on Visual Analytics Science and Technology VAST’08, pp.19-26
Garg, S., Ramakrishnan, I.V., Mueller, K. A 2010. Visual Analytics Approach to Model Learning.
In Proc. IEEE Symposium on Visual Analytics Science and Technology VAST’10, pp.67-74
Guo, D. 2009. Multivariate Spatial Clustering and Geovisualization. In Geographic Data Mining
and Knowledge Discovery, edited by H. J. Miller and J. Han. London and New York, Taylor &
Francis, pp. 325-345.
Guo,Z., Ward, M.O., Rundensteiner, E.A. 2009. Model Space Visualization for Multivariate
Linear Trend Discovery. In Proc. IEEE Symposium on Visual Analytics Science and Technology
VAST’09, pp.75-82
Hao, M. C., Janetzko, H., Mittelstädt, S., Hill, W., Dayal, U., Keim, D. A., Marwah, M., Sharma,
R. K. 2011. A Visual Analytics Approach for Peak-Preserving Prediction of Large Seasonal Time
Series. Computer Graphics Forum, 30 (3), pp. 691-700.
Kamarianakis, Y., Prastacos, P. 2003. Forecasting Traffic Flow Conditions in an Urban Network:
Comparison of Multivariate and Univariate Approaches. Transportation Research Record-Journal
of the Transportation Research Board, No. 1857, pp. 74-84.
Kamarianakis, Y., Prastacos, P. 2005. Space - Time Modeling of Traffic Flow. Computers and
Geosciences, 31, pp. 119-133
Kamarianakis, Y., Prastacos, P. 2006. Spatial Time-Series Modeling: A review of the proposed
methodologies. Working Papers of the University of Crete, Department of Economics, No. 0604,
http://ideas.repec.org/p/crt/wpaper/0604.html, last accessed September 19, 2011
Keim, D., Andrienko, G., Fekete, J.-D., Görg, C., Kohlhammer, J., Melançon, G. 2008. Visual
Analytics: Definition, Process, and Challenges. In: Kerren, A., Stasko, J.T., Fekete, J.-D., North,
C. (editors). Information Visualization – Human-Centered Issues and Perspectives. Lecture Notes
in Computer Science, Vol. 4950, Springer, Berlin, pp.154-175
Kohonen, T. 2001. Self-Organizing Maps. Springer, Berlin
Kyriakidis, P.C., Journel, A.G. 2001. Stochastic modeling of atmospheric pollution: a spatial
time-series framework. Part I: methodology; Part II: application to monitoring monthly sulfate
deposition over Europe. Atmospheric Environment, 35, pp. 2331-2337; 2339-2348.
Maciejewski, R., Rudolph, S., Hafen, R., Abusalah, A., Yakout, M., Ouzzani, M., Cleveland,
W.S., Grannis, S.J., Ebert, D.S. 2010. A Visual Analytics Approach to Understanding
Spatiotemporal Hotspots. IEEE Transactions on Visualization and Computer Graphics. 16(2),
pp.205-220.
-
36
Maciejewski, R., Livengood, P., Rudolph, S., Collins, T.F., Ebert, D.S., Brigantic, R.T., Corley,
C.D., Muller, G.A., Sanders, S.W. 2011. A Pandemic Influenza Modeling and Visualization Tool.
Journal of Visual Languages and Computing, 22, pp.268-278
Matković, K., Gračanin, D., Jelović, M., Ammer, A., Lež, A., Hauser, H. 2010. Interactive Visual
Analysis of Multiple Simulation Runs Using the Simulation Model View: Understanding and
Tuning of an Electronic Unit Injector, IEEE Transactions on Visualization and Computer
Graphics, 16(6), pp.1449-1457
Matković, K., Gračanin, D., Jelović, M., and Cao, Y. 2011. Adaptive Interactive Multi-Resolution
Computational Steering for Complex Engineering Systems. In Proceedings of EuroVA, Bergen,
Norway, pp. 45-48
Migut, M., Worring, M. 2010. Visual Exploration of Classification Models for Risk Assessment.
In Proc. IEEE Symposium on Visual Analytics Science and Technology VAST’10, pp. 11-18
Rinzivillo, S., Pedreschi, D., Nanni, M., Giannotti, F., Andrienko, N., and Andrienko, G. Visually–
driven analysis of movement data by progressive clustering, Information Visualization, 7(3/4),
2008, 225-239.
Sammon, J. W. 1969. A nonlinear mapping for data structure analysis. IEEE Transactions on
Computers, 18, pp.401–409
Schreck, T., Bernard, J., von Landesberger, T., Kohlhammer, J. 2009. Visual cluster analysis of
trajectory data with interactive Kohonen maps. Information Visualization, 8(1), pp.14–29
Slingsby, A., Wood, J., Dykes, J. Clouston, D., Foote, M., 2010. Visual analysis of sensitivity in
CAT models: interactive visualisation for CAT model sensitivity analysis, In Proceedings of
Accuracy 2010 Conference, Leicester, UK, 20-23 July 2010.
Therón, R., De Paz, J.F. 2006. Visual Sensitivity Analysis for Artificial Neural Networks In:
Lecture Notes in Computer Science. IDEAL 2006, vol. 4224, pp. 191-198, Berlin: Springer
Xiao, L., Gerth, J., Hanrahan, P. 2006. Enhancing Visual Analysis of Network Traffic Using a
Knowledge Representation. In Proc. IEEE Symposium on Visual Analytics Science and
Technology VAST’06, pp. 107-114
Ziegler, H., Jenny, M., Gruse, T., Keim, D.A. 2010. Visual market sector analysis for financial
time series data. In Proc. IEEE Symposium on Visual Analytics Science and Technology
VAST’10, pp.83-90